Historical AI Training Data

The internet
before the
internet

33 years of authentic human discourse — before SEO, before algorithms, before AI. The Usenet Corpus 1980–2013 is one of the largest privately held pre-LLM text datasets in existence.

103.1B Tokens (cl100k_base)
408M Posts
18,347 Newsgroups
33 yrs 1980 – 2013
0 AI-generated posts

Nine hierarchies.
Three decades.

Every major Usenet newsgroup hierarchy, processed and validated. From technical discourse that built the internet to cultural conversations that defined an era.

alt.*
58.8B tokens
228,761,734 posts · 15,288 groups
The largest and most diverse hierarchy. Culture, philosophy, hobbies, and everything in between.
rec.*
16.5B tokens
77,233,777 posts · 919 groups
Recreation, arts, sports, and hobbies. 99.7% English, exceptionally clean signal.
comp.*
10.3B tokens
49,990,816 posts · 1,205 groups
Technical computing discourse from the people who built the internet. Rare early-era coverage from 1980.
soc.*
8.2B tokens
22,386,509 posts · 341 groups
Social topics and culture. Most multilingual hierarchy — 17.2% non-English from soc.culture.* groups.
sci.*
3.3B tokens
11,996,714 posts · 237 groups
Scientific research and debate. Dense technical content across physics, biology, mathematics, and more.
misc.*
2.7B tokens
10,866,446 posts · 242 groups
Jobs, kids, health, and everyday life. High relevance for social science and sentiment research.
news.*
1.7B tokens
2,065,102 posts · 60 groups
Usenet meta-discussion and announcements. Highest average post length — includes compiled FAQ documents.
talk.*
1.3B tokens
4,308,542 posts · 47 groups
Debate and argumentation. Politics, religion, philosophy. Rich adversarial discourse.
humanities.*
0.2B tokens
626,648 posts · 8 groups
Literature, philosophy, classics. Smaller but high signal — 99.6% English.
Token volume by year — 1980 to 2013
1980199020002013

Built for how
researchers actually work.

License the full corpus or a targeted subset. Academic and commercial terms available. All licenses include delivery via private HuggingFace repository.

Big 8
Curated Corpus
~44B tokens · exc. alt.*

The eight major hierarchies without alt.*. Higher signal-to-noise ratio for research requiring cleaner, more focused training data.

  • comp, rec, soc, sci, misc, news, talk, humanities
  • Reduced noise from alt.* excluded
  • Ideal for domain-specific fine-tuning
  • Academic and commercial licenses available
Request proposal →
Custom Subset
Targeted Extract
Your specification

Specific hierarchies, date ranges, or topic areas. Previous subsets include pre-2005 temporal slices and sentiment-focused social group extracts.

  • Any hierarchy or combination
  • Any date range 1980–2013
  • Token-count verified before delivery
  • Academic pricing for qualified researchers
Request proposal →
Coming Soon
Threaded Corpus
Full conversation trees

Thread-reconstructed version with conversation chain GUIDs, thread position, and depth fields. Rebuilt from original MBOX source files with References headers intact.

  • Full conversation threading preserved
  • thread_id, thread_position, thread_depth fields
  • Ideal for dialogue and reasoning model training
  • Contact us to be notified at release
Join waitlist →

Who uses
this corpus.

Active research and licensing inquiries span AI labs, economics research, computational linguistics, and digital humanities.

LLM Pretraining

Language Model Training

Pre-web, human-only discourse with zero AI contamination. Covers language evolution across three decades before engagement optimization existed.

Chronological Modeling

Temporal LLMs

Timestamped at post level across 33 years. Ideal for chronologically consistent models that need text data available only at specific points in time.

Social Science

Economic & Social Research

Large-scale text analysis of sentiment, attitude, and discourse evolution. Currently in use for generational economic sentiment research at leading universities.

Domain Fine-tuning

Specialized Adaptation

The comp.* and sci.* hierarchies offer dense technical discourse for domain-specific fine-tuning. Early internet computing culture is underrepresented in modern datasets.

Historical Preservation

Digital Humanities

A continuous record of public discourse from networked academic communities through mass internet adoption. An irreplaceable window into pre-social-media culture.

Dialogue Research

Conversational AI

Threaded corpus version in development will provide full conversation trees — multi-turn argumentation and debate at scale, from authentic human exchanges.

How it was built.

A two-stage cleaning pipeline applied to raw MBOX archives. 99% of files validated clean with zero email or PII violations.

  • 01

    Hierarchy-level exclusions

    alt.binaries.* removed entirely — UUencoded binary content with negligible NLP value. Adult content newsgroups excluded. Most binary data removed before record-level processing began.

  • 02

    Deduplication

    Posts with identical Message-IDs removed. Cross-file deduplication applied. Usenet cross-posting artifacts eliminated.

  • 03

    Binary detection

    MIME header inspection for residual UUencoded and base64 binary content. Records with binary payloads dropped entirely.

  • 04

    PII redaction

    Email addresses replaced with [email] token throughout post bodies. Author email addresses stripped. Message-IDs SHA-256 hashed to msg-<hex32>.

  • 05

    Format conversion

    MBOX → structured JSONL, gzip-compressed. Language detection on every record using Meta's fasttext LID-176. 96.6% English, 100+ languages total.

  • 06

    Validation

    Full-corpus validation scan with zero-tolerance checks for email addresses, binary content, and duplicate IDs. 99% of 18,347 files validated clean.

Start a licensing conversation.

Academic and commercial licenses available. Custom subsets, temporal slices, and hierarchy-specific extracts on request. Typical response within 2 business days.

todd@ownedbydanes.com